An FPGA Based Heterogeneous Accelerator for Single Shot MultiBox Detector (SSD)
Liang Cai, Feng Dong, Ke Chen, Kehua Yu, Wei Qu, Jianfei Jiang · 2020
Single Shot MultiBox Detector (SSD) is a commonly used neural networks for real-time object detection, which has high requirements for inference speed. In this paper, an accelerator on a heterogeneous CPU-FPGA platform is proposed. An optimal graph partition method is also presented. By doing optimizations on both software and hardware sides, considerable performance improvement has been observed. Compared with CPU, experiment results demonstrate that the proposed accelerator achieves 4.5x acceleration performance for SSD_Mobilenet_V2. For SSD_Inception_V2 model, our accelerator achieves 1.42x that of Nvidia Titan X GPU.